@dakotabeat Code on github but you will have to compile the app yourself for now, we'll try to make an actual app in the coming weeks.

RT @kyutai_labs: Thanks to @Xavier75 for stopping by at the #AIActionSummit to try Hibiki. No need to struggle with English anymore 😅 https…

@fffiloni @kyutai_labs @Xavier75 Amazing, thanks for putting this up! Also looks like hibiki works well at translating French songs🎵 though it doesn't really sing so far 😆

RT @fffiloni: Here one @kyutai_labs speech-to-speech Hibiki example with @Xavier75 🤗

Afraid of missing out on French pop culture references because you don't speak the language? Fear no more and try our Hibiki speech-to-speech translation model— no more FOMO! 🇫🇷✨ #Translation #AI

@mstefanec @kyutai_labs Yes supporting more languages is certainly on our todo list. Supporting new languages mostly requires generating some synthetic data. As we've published all the details in our tech report, we also hope other groups can work on adding different languages.

@fdaudens @kyutai_labs Thanks :)

@awnihannun Thanks @awnihannun , your help with MLX was really helpful to make this happen!

Neil showing Hibiki 🟢 running on an iphone. All local so whatever you say will stay on the iphone and latency is great! Get the code to build the ios app in our moshi-swift repo (it's brittle at the moment but we'll polish it 😅 )

@zackangelo Yes we do have a rust inference that is candle based, I'll put some code snippet to show how to run it soonish!

Very impressive to hear this Japanese 🇯🇵 version of moshi 🟢. I don't speak the language so I cannot understand what it's trying to tell me but at least it sounds great 😅

@adelmoumen_ Jax et PyTorch ont leurs tradeoffs et aucun ne me paraît clairement superieur. L'idéal est que chaque équipe puisse travailler avec la techno qu'elle préfère et ça nous permet de mieux connaître les pros et cons de chaque côté.

@adelmoumen_ Jax est utilisé pour la partie text only, PyTorch pour les modèles multimodaux/audios. Dans les deux cas, le dataloader est fait in-house principalement en rust.

@ghorbani_asghar @kyutai_labs Ah actually when the phone is plugged in, helium 2b also reaches ~35 tok/s (just saw a 36 tok/s) - not sure which of the two numbers should be reported. Certainly hoping that the community gets a gguf export soon and people start hacking with it!

@kyutai_labs And the inference speed even reaches 36 tok/s when the phone is plugged in 🚀

RT @kyutai_labs: Helium 2B running locally on an iPhone 16 Pro at 28 tok/s, faster than you can read, all thanks to mlx-swift with q4 quant…

RT @neilzegh: I’m honored that SoundStream (2021) retrospectively receives the best paper award from @IEEEsps. We introduced neural audio…